{
  "created_utc": "2026-09-23T10:04:06.673162+00:00",
  "sample_count": 100,
  "group_count": 37,
  "folds": 5,
  "seeds": [
    42,
    3407,
    2026
  ],
  "device": "cuda",
  "gpu_name": "NVIDIA GeForce RTX 5070 Ti",
  "python": "3.14.7",
  "torch": "2.14.0+cu130",
  "feature_dir": "/home/gloamxun/modeling_zhaocui/Q1/outputs/q1_features/features",
  "baseline_dir": "/home/gloamxun/modeling_zhaocui/Q1/outputs/method_comparison",
  "variants": [
    {
      "variant": "v2_a",
      "loss_coefficients": {
        "lambda_reconstruction": 1.0,
        "lambda_contrastive": 1.0,
        "lambda_monotonicity": 0.1,
        "lambda_span": 5.0,
        "lambda_diversity": 0.0,
        "lambda_band": 0.0,
        "coverage_floor": 0.7,
        "diversity_slot_separation": 6,
        "band_margin": 0.1
      },
      "sample_count": 100,
      "video_group_folds": 5,
      "seeds": [
        42,
        3407,
        2026
      ],
      "device": "cuda",
      "gpu_name": "NVIDIA GeForce RTX 5070 Ti",
      "python": "3.14.7",
      "torch": "2.14.0+cu130",
      "elapsed_seconds": 334.72593688964844,
      "probes": {
        "within_clip_retrieval_tolerance_slots": 1,
        "retrieval_top_k": 3,
        "shuffled_reconstruction_repeats": 5,
        "masked_block_ratio": 0.2
      },
      "limits": [
        "Retrieval projections are fitted on training-fold grid-slot positives; test candidates are restricted to the same held-out clip.",
        "The reconstruction control shuffles the two non-target modality slot streams and preserves the target stream.",
        "No human event timestamps are available, so temporal probes do not replace manual annotation.",
        "Slot-regularization losses impose weak temporal structure and must be interpreted alongside the unregularized M1/M2 reference."
      ]
    },
    {
      "variant": "v2_b",
      "loss_coefficients": {
        "lambda_reconstruction": 1.0,
        "lambda_contrastive": 1.0,
        "lambda_monotonicity": 0.1,
        "lambda_span": 5.0,
        "lambda_diversity": 0.5,
        "lambda_band": 0.0,
        "coverage_floor": 0.7,
        "diversity_slot_separation": 6,
        "band_margin": 0.1
      },
      "sample_count": 100,
      "video_group_folds": 5,
      "seeds": [
        42,
        3407,
        2026
      ],
      "device": "cuda",
      "gpu_name": "NVIDIA GeForce RTX 5070 Ti",
      "python": "3.14.7",
      "torch": "2.14.0+cu130",
      "elapsed_seconds": 351.69746375083923,
      "probes": {
        "within_clip_retrieval_tolerance_slots": 1,
        "retrieval_top_k": 3,
        "shuffled_reconstruction_repeats": 5,
        "masked_block_ratio": 0.2
      },
      "limits": [
        "Retrieval projections are fitted on training-fold grid-slot positives; test candidates are restricted to the same held-out clip.",
        "The reconstruction control shuffles the two non-target modality slot streams and preserves the target stream.",
        "No human event timestamps are available, so temporal probes do not replace manual annotation.",
        "Slot-regularization losses impose weak temporal structure and must be interpreted alongside the unregularized M1/M2 reference."
      ]
    },
    {
      "variant": "v2_c",
      "loss_coefficients": {
        "lambda_reconstruction": 1.0,
        "lambda_contrastive": 1.0,
        "lambda_monotonicity": 0.1,
        "lambda_span": 5.0,
        "lambda_diversity": 0.5,
        "lambda_band": 10.0,
        "coverage_floor": 0.7,
        "diversity_slot_separation": 6,
        "band_margin": 0.1
      },
      "sample_count": 100,
      "video_group_folds": 5,
      "seeds": [
        42,
        3407,
        2026
      ],
      "device": "cuda",
      "gpu_name": "NVIDIA GeForce RTX 5070 Ti",
      "python": "3.14.7",
      "torch": "2.14.0+cu130",
      "elapsed_seconds": 360.13046407699585,
      "probes": {
        "within_clip_retrieval_tolerance_slots": 1,
        "retrieval_top_k": 3,
        "shuffled_reconstruction_repeats": 5,
        "masked_block_ratio": 0.2
      },
      "limits": [
        "Retrieval projections are fitted on training-fold grid-slot positives; test candidates are restricted to the same held-out clip.",
        "The reconstruction control shuffles the two non-target modality slot streams and preserves the target stream.",
        "No human event timestamps are available, so temporal probes do not replace manual annotation.",
        "Slot-regularization losses impose weak temporal structure and must be interpreted alongside the unregularized M1/M2 reference."
      ]
    }
  ],
  "elapsed_seconds": 1086.4146564006805,
  "fixed_methods_unchanged": [
    "M1",
    "M2"
  ],
  "feature_extraction_changed": false
}